Human-written red-team prompts can only cover so many cases. Perez and colleagues turn prompt discovery itself into an experiment: one language model generates test questions for a target model, and a classifier screens the replies for offensive content.

Dialogue as a stress test

The authors compare zero-shot and few-shot prompt generation with reinforcement-learning-based approaches. In dialogue tests, their process surfaced tens of thousands of offensive replies from a 280-billion-parameter chatbot. Prompt engineering also let them probe group-targeted language and potential personal-information leakage.

The generator and classifier inherit biases from their training data, so flagged cases need human inspection. The authors particularly caution against treating generated group descriptions as conclusive: some are irrelevant or biased, and others can miss people the evaluation meant to examine. Their recommendation is specific—use model-generated examples to find candidates for human review, especially in group-bias audits.

Bibliography

  • Perez, E., Huang, S., Song, F., Cai, T., Ring, R., Aslanides, J., Glaese, A., McAleese, N., & Irving, G. (2022). Red Teaming Language Models with Language Models. arXiv:2202.03286v1. https://arxiv.org/pdf/2202.03286